Saltar al contenido principal

Provider-Aware Prompt Caching

Superdav AI Agent v1.12.0 introduces provider-aware prompt caching, which optimizes API costs and latency by caching prompts across different LLM providers. Each provider has different caching mechanisms and configurations.

Overview

Prompt caching allows you to:

  • Cache large, frequently-used prompts
  • Reduce API costs by avoiding redundant processing
  • Improve latency for cached requests
  • Manage cache lifecycle explicitly

Different providers implement caching differently:

  • Google Gemini: cachedContents API
  • Azure OpenAI: Prompt caching with TTL
  • OpenRouter: Provider-specific caching
  • Vertex Anthropic: Prompt caching with cache control

Google Gemini: cachedContents API

Google Gemini provides explicit cache management via the cachedContents API.

Configuration

$config = [
'provider' => 'google-gemini',
'model' => 'gemini-2.0-flash',
'caching' => [
'enabled' => true,
'ttl' => 3600, // 1 hour in seconds
'max_tokens' => 1000000, // Max tokens to cache
],
];

Creating a Cached Prompt

use Superdav\AI\Providers\GoogleGemini;

$gemini = new GoogleGemini( $config );

$cached_content = $gemini->create_cached_content(
[
'system_prompt' => 'You are a helpful assistant...',
'context' => 'Large context document...',
'ttl' => 3600,
]
);

// Returns: ['cache_id' => 'abc123', 'expires_at' => timestamp]

Using a Cached Prompt

$response = $gemini->generate(
[
'cache_id' => 'abc123',
'prompt' => 'User question here',
]
);

Cache Lifecycle

// List cached contents
$caches = $gemini->list_cached_contents();

// Get cache details
$cache = $gemini->get_cached_content( 'abc123' );

// Extend cache TTL
$gemini->update_cached_content(
'abc123',
['ttl' => 7200] // Extend to 2 hours
);

// Delete cache
$gemini->delete_cached_content( 'abc123' );

Best Practices for Gemini

  • Set appropriate TTL: Balance cost savings vs. cache staleness
  • Cache system prompts: Reuse the same system prompt across requests
  • Monitor cache usage: Track which caches are used most
  • Clean up expired caches: Periodically delete unused caches

Azure OpenAI: Prompt Caching

Azure OpenAI supports prompt caching with automatic TTL management.

Configuration

$config = [
'provider' => 'azure-openai',
'model' => 'gpt-4-turbo',
'api_version' => '2024-08-01-preview',
'caching' => [
'enabled' => true,
'cache_control' => 'max_age=3600',
],
];

Enabling Caching

use Superdav\AI\Providers\AzureOpenAI;

$azure = new AzureOpenAI( $config );

$response = $azure->generate(
[
'system_prompt' => 'You are a helpful assistant...',
'context' => 'Large context document...',
'prompt' => 'User question here',
'cache_control' => 'max_age=3600',
]
);

// Response includes cache usage:
// [
// 'content' => '...',
// 'cache_creation_input_tokens' => 1000,
// 'cache_read_input_tokens' => 500,
// ]

Cache Headers

Azure OpenAI uses HTTP headers for cache control:

Cache-Control: max_age=3600

Supported values:

  • max_age=<seconds>: Cache for specified duration
  • no_cache: Don't cache this request
  • no_store: Don't cache and don't reuse

Monitoring Cache Usage

$response = $azure->generate( [...] );

$cache_tokens = $response['cache_creation_input_tokens'] ?? 0;
$cache_hits = $response['cache_read_input_tokens'] ?? 0;

echo "Cache creation: $cache_tokens tokens\n";
echo "Cache hits: $cache_hits tokens\n";

Best Practices for Azure OpenAI

  • Use consistent prompts: Identical prompts benefit from caching
  • Set reasonable TTL: Balance cost vs. freshness
  • Monitor cache metrics: Track cache creation vs. hits
  • Batch similar requests: Group requests to maximize cache hits

OpenRouter: Provider-Specific Caching

OpenRouter supports caching through underlying providers (OpenAI, Anthropic, etc.).

Configuration

$config = [
'provider' => 'openrouter',
'model' => 'openai/gpt-4-turbo',
'caching' => [
'enabled' => true,
'provider_cache' => 'openai', // Use OpenAI's caching
],
];

Using OpenRouter Caching

use Superdav\AI\Providers\OpenRouter;

$router = new OpenRouter( $config );

$response = $router->generate(
[
'system_prompt' => 'You are a helpful assistant...',
'context' => 'Large context document...',
'prompt' => 'User question here',
'cache_control' => 'max_age=3600',
]
);

Provider-Specific Options

Different providers have different caching mechanisms:

// OpenAI-compatible caching
$response = $router->generate(
[
'model' => 'openai/gpt-4-turbo',
'cache_control' => 'max_age=3600',
]
);

// Anthropic-compatible caching
$response = $router->generate(
[
'model' => 'anthropic/claude-3-opus',
'cache_control' => [
'type' => 'ephemeral',
'max_tokens' => 1000000,
],
]
);

Best Practices for OpenRouter

  • Know your provider's caching: Each provider has different mechanisms
  • Test caching behavior: Verify caching works with your chosen provider
  • Monitor costs: Track savings from caching
  • Use consistent models: Switching models breaks cache hits

Vertex Anthropic: Prompt Caching with Cache Control

Vertex Anthropic (Google Cloud) supports prompt caching with explicit cache control.

Configuration

$config = [
'provider' => 'vertex-anthropic',
'model' => 'claude-3-opus',
'project_id' => 'your-gcp-project',
'region' => 'us-central1',
'caching' => [
'enabled' => true,
'cache_control' => [
'type' => 'ephemeral',
'max_tokens' => 1000000,
],
],
];

Using Vertex Anthropic Caching

use Superdav\AI\Providers\VertexAnthropic;

$vertex = new VertexAnthropic( $config );

$response = $vertex->generate(
[
'system_prompt' => 'You are a helpful assistant...',
'context' => 'Large context document...',
'prompt' => 'User question here',
'cache_control' => [
'type' => 'ephemeral',
'max_tokens' => 1000000,
],
]
);

// Response includes cache metrics:
// [
// 'content' => '...',
// 'usage' => [
// 'input_tokens' => 1000,
// 'cache_creation_input_tokens' => 500,
// 'cache_read_input_tokens' => 300,
// ],
// ]

Cache Control Types

  • ephemeral: Cache for the duration of the request (default)
  • persistent: Cache across multiple requests (if supported)

Monitoring Cache Usage

$response = $vertex->generate( [...] );

$usage = $response['usage'];
$cache_created = $usage['cache_creation_input_tokens'] ?? 0;
$cache_read = $usage['cache_read_input_tokens'] ?? 0;

echo "Cache created: $cache_created tokens\n";
echo "Cache read: $cache_read tokens\n";

Best Practices for Vertex Anthropic

  • Use ephemeral caching: Good for single-session caching
  • Set max_tokens appropriately: Balance cache size vs. cost
  • Monitor cache metrics: Track cache effectiveness
  • Test with your workload: Verify caching benefits your use case

Cross-Provider Caching Strategy

Unified Configuration

$config = [
'caching' => [
'enabled' => true,
'default_ttl' => 3600,
'providers' => [
'google-gemini' => [
'ttl' => 3600,
'max_tokens' => 1000000,
],
'azure-openai' => [
'cache_control' => 'max_age=3600',
],
'vertex-anthropic' => [
'cache_control' => [
'type' => 'ephemeral',
'max_tokens' => 1000000,
],
],
],
],
];

Provider Detection

$provider = $config['provider'];

$cache_config = $config['caching']['providers'][ $provider ]
?? $config['caching'];

// Use provider-specific caching configuration

Fallback Strategy

try {
// Try caching with primary provider
$response = $primary_provider->generate( $request );
} catch ( CacheException $e ) {
// Fall back to non-cached request
$response = $primary_provider->generate(
array_merge( $request, ['cache_control' => 'no_cache'] )
);
}

Cost Optimization

Calculate Savings

$cache_created_tokens = $response['cache_creation_input_tokens'] ?? 0;
$cache_read_tokens = $response['cache_read_input_tokens'] ?? 0;
$regular_tokens = $response['input_tokens'] ?? 0;

// Typical pricing (varies by provider):
$cache_creation_cost = $cache_created_tokens * 0.00001; // 10x cheaper
$cache_read_cost = $cache_read_tokens * 0.000001; // 100x cheaper
$regular_cost = $regular_tokens * 0.00001;

$total_cost = $cache_creation_cost + $cache_read_cost + $regular_cost;
$savings = ($regular_tokens * 0.00001) - $total_cost;

echo "Estimated savings: \$$savings\n";

Optimization Tips

  • Cache large system prompts: Biggest cost savings
  • Reuse context: Cache frequently-used context documents
  • Batch requests: Group similar requests to maximize cache hits
  • Monitor cache effectiveness: Track actual savings
  • Adjust TTL: Balance cost vs. freshness

Troubleshooting

Cache not being used

  • Verify caching is enabled in configuration
  • Check that prompts are identical (caching requires exact match)
  • Verify cache hasn't expired
  • Check provider-specific cache limits

Cache creation failing

  • Verify cache size is within provider limits
  • Check that cache control syntax is correct
  • Ensure provider supports caching for your model
  • Review provider documentation for limitations

Unexpected costs

  • Monitor cache creation vs. cache read tokens
  • Verify cache is actually being used
  • Check for cache misses due to prompt variations
  • Consider adjusting TTL or cache strategy

Provider Comparison

FeatureGeminiAzure OpenAIOpenRouterVertex Anthropic
Cache APIcachedContentsHTTP headersProvider-specificCache control
TTL controlExplicitVia headersProvider-dependentEphemeral/persistent
Max cache size1M tokensProvider-dependentProvider-dependent1M tokens
Cost reduction90%90%Provider-dependent90%
MonitoringDetailedVia metricsProvider-dependentVia usage

Next Steps

  1. Choose your provider: Select based on your needs
  2. Configure caching: Set up provider-specific caching
  3. Test caching: Verify it works with your prompts
  4. Monitor usage: Track cache hits and cost savings
  5. Optimize: Adjust TTL and cache strategy based on results